George

@KuittinenPetri @antirez Is the Beelink one of the boxes that claims 80gbps usb4 ports? have you tested that? thanks

I have both ASUS Ascent GX10 (Nvidia GB10 clone) and Beelink GTR9 Pro AMD Ryzen™ AI Max+ 395. Both have 128 GB LPDDR5X. Nvidia GB10 is slightly faster in decode, thanks to 273 GB/s memory bandwidth vs ~240 GB/s for AMD Strix Halo. I like both them and here I briefly explain why: + both have excellent thermals (Beelink is probably the best among Strix Halo boxes in this respect, never gets overheated, massive heat sink, fan does make some noise, also ASUS is the heaviest Nvidia GB10 clone, implying the largest full copper heat sink) + AMD Ryzen™ AI Max+ 395 is surprisingly good in gaming, able to run older games 4k resolution at high details and one can still run AI loads at same time (!), temperature stays +65 C and game doesn't even slow down, but gives 60 Hz (my display has only 60 Hz) + AMD is ~12% slower in decode, due slower memory bandwidth + AMD of course loses in prefill vs Nvidia GBA10, but it can still reach ~500 token/s with some models + AMD allows only max 96 GB of RAM used as VRAM / AI loads + ADM NPU is not used my llama.cpp or most other software, everything is done with GPU. I don't know why they even put in 50 AI TOPS NPU, if the software won't support it + AMD's 16-core CPU is very fast with that 128 GB fast RAM. Windows boots and runs extremely smoothly without any hiccups. This could be important in many CPU-heavy agentic loads. + Nvidia has better software compatibility and better optimized, AMD could perhaps in some cases get very large gains with better optimization

I guess I should wait a few weeks and buy a AMD Ryzen AI Halo Box, instead of the current alternatives? I want ROCm support in DS4 to be good and well supported.

claude is in the bios

uh oh subsidybros, did we get too cocky?

@ErgodicTrace me too- seems to be especially common in our generation of the 'ai at home' crowd. it's like something from classical tragedy- having failed to to find life in the heavens, we set about creating it ourselves

folding@home for demis, i remember using my parents electricity running my ps3 24/7 to do this. the cell processor was a monster once developers realized how to optimize the binaries for the spes (synergistic processing elements). developers mastered it in the late sustainment cycle for rendering and physics.

@phoronix memory bandwidth pls

RISC-V RVA23 hardware just arrived! K3 in the lab! Benchmarks ahead.

so you're telling me tanstack is nothing to do with garry tan? gstack bros stay winning

if you die in initrd, you die in real life

i think i prefer codex's- classic torus knot, tasteful tunnel effect, chefs kiss

free alpha for ai-assisted embedded dev- uhubctl- lets you programmatically power on/off usb ports so your agent can reset wedged device when you leave it on full-auto. ridiculously powerful. left it in full control of /dev/ttyACM0 and asked for a demoscene-style animation:

RISCV Hellas anyone? NPU might take some time to get working (buy a strix halo) but client already "just works" via nixos cross-compile, eternal thanks to maintainers :)

running `cargo clean` after a long time feels so good, like taking a huge dump

@IterIntellectus playing a worse video game

what’s the female equivalent of a man playing video games?

end result: memory-safe programming languages exist, encryption used where appropriate; firewalls and sandboxes exist. the alternative scenario (which ai-safety-ists promote): 'grep' is export-controlled, only legal for qualified greppers to use; mention of 40yo bugs forbidden

@jawwwn_ as a maligned tech person, i really hope he's right

Palantir CEO Alex Karp rips into AI slop companies: “The appearance of software working, is not software working.” “The slop getting a lot of attention is dangerous in terms of the hyperbolic rhetoric—that there will be no jobs because of slop, nothing will work, we will have a god-like figure in the name of AI.” “In fact, what actually does work is a platform built by a motley crew of highly-technical people who over 20 years have been maligned for being right.”

i hope you're ready, anon

there shall be abliterated research models for the cybersecurity researchers; porno LoRas for the perverts, furry finetunes for the furries. just like the internet- the bottomless well; the human soul.

One of the things I miss the most about living in SF is meeting founders of products i use and describing how i abuse their product for purposes other than the intended

TIL on "Railway" if you delete a resource id in staging you delete it in prod. ycslop?

@sama @TheAhmadOsman how is democracy compatible with gated access? laughable. release the weights!

@TheAhmadOsman War is peace. Freedom is slavery. Ignorance is strength. oh wait, we don't believe any of that. how about we democratize a lot of super capable AI, and then we sit back and watch you build the future?

The difference between Anthropic and OpenAI is that one of them consistently keeps gaslighting us about not being an evil company Big brother energy in the worst possible way